Fix quark fp8 format loading.

[fxmarty-amd]

This PR fixes an issue in the order requantize_with_max_scale is called when loading a checkpoint using quark config.json/checkpoint format.

Compare the following:

• neuralmagic fp8: https://github.com/vllm-project/vllm/blob/7a8987dac5f0ed0c798a73e8b4ec8f5e640bc63a/vllm/model_executor/layers/quantization/fp8.py#L303-L317
• quark fp8:

  vllm/vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py

  Lines 35 to 48 in 6852819

  	if self.qscheme == "per_tensor":
  	max_w_scale, weight = requantize_with_max_scale(
  	weight=layer.weight,
  	weight_scale=layer.weight_scale,
  	logical_widths=layer.logical_widths,
  	)
  	if current_platform.is_rocm():
  	weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
  	weight=weight,
  	weight_scale=max_w_scale,
  	input_scale=layer.input_scale)
  	if input_scale is not None:
  	layer.input_scale = Parameter(input_scale,

normalize_e4m3fn_to_e4m3fnuz should be called before requantize_with_max_scale.

[gshtras]

A separate PR is not required. Once the upstream one is accepted, we'll get it within 1 week usually

[fxmarty-amd]

@gshtras I think ROCm/vllm is used more widespread internally so if this can get included here before merging upstream, it is nice, but I can close this if you prefer!

[BowenBao]

both weight and max_w_scale are undefined if current_platform.is_rocm() is False.

[BowenBao]

Do you observe better accuracy with this change? Mathematically the order should not matter.

edit: taking this back, the order does matter since otherwise the requantize_with_max_scale is requanting into fnuz using fn scales. Let's point this out in the PR description.